Lessons learned from robotics grasping and manipulation competition

Two weeks ago, I, along with eight researchers in the IEEE RAS TC on Robotic Hands, Grasping, and Manipulation, organized the first Robotics Grasping and Manipulation Competition in Daejeon, South Korea during IROS 2016. The competition had three tracks: hand-in-hand, fully autonomous (only in grasping stage), and simulation. There are daily-living manipulation tasks and pick-anyplace tasks. Totally nine teams participated. Some teams participated in multiple tracks. The competition’s web page is at
http://www.rhgm.org/activities/competition_iros2016/ . I have wrote a summary in ppt and its pdf file is at
http://www.rhgm.org/activities/competition_iros2016/competition_iros_summary.pdf .

Here are my key thoughts on the competition.
1. The robotic hands used in the competition are very good. Many of them were able to do pick-and-place tasks easily even in quite difficult situations.

2. Human intelligence can compensate some of the design flaws in the robotic hands for complicated manipulation tasks. For example, using a pair of scissors was very challenging for all robotic hands. However, the human operators were able to find effective tricks to overcome the lack of dexterity in the robotic hands and finish the tasks, but use much more time than using our hands.

3. All teams preferred simple grasps: power grasp when possible. Grasping plan were done mostly through trial and error during the dry-run. Efficiency of manipulation in full auto track were barely considered by any team.

4. If we assume the robots had perfect visual perception (as good as our vision), the robots in the competition could finish about 80% of the manipulation tasks automatically slowly. The rest 20% are too complicated or requires too much power from the robot.

5. With imperfect visual perception, the robots in the competition had trouble to use the perception to figure out how to escape from a difficult situation such as large object blocks other objects, where the large object is too heave to be picked out first.

6. For pick-and-place tasks, computer vision dominates the problems. Robotic hand design plays very small roll here. Even relatively simple grippers would work fine in hand-in-hand track, but works poorly in full-auto with robot vision.

7. Perfect vision is very helpful in manipulation. However, the physical interaction in the manipulation requires a lot more. Both the hardware and the AI need to improve dramatically even with perfect computer vision.

8. Many great challenges for robotics lay in the physical interactive manipulation tasks.


Leave a Reply

Your email address will not be published. Required fields are marked *